Google DeepMind Introduces AI Agent SIMA: AI That Can Beat Any Game
Google DeepMind introduced SIMA, an AI agent that can successfully complete any video game. Experts emphasized that artificial intelligence learns to play as if it were a gamer, not a robot.
To train and test SIMA, the DeepMind lab collaborated with eight game developers, including Hello Games, Embracer, Tuxedo Labs, Coffee Stain. The researchers connected SIMA to games such as No Man’s Sky, Teardown, Valheim and Goat Simulator 3. The developers claim that SIMA does not require a special API to run games or access to source code.
SIMA combines natural language learning with understanding 3D worlds and image recognition. “SIMA is not trained to win a game, it is trained to run a game and do what it is told,” says Google DeepMind researcher Tim Harley. The team chose games that were more open-world focused than narrative-driven to help SIMA learn general gaming skills. For example, the essence of the Goat Simulator game is to perform random, spontaneous actions, and this is exactly the kind of spontaneity that the developers wanted to achieve from their AI agent.
SIMA currently has about 600 basic skills, such as turning left, climbing stairs, and opening a menu to use a map. SIMA could eventually be tasked with performing more complex functions in the game. Tasks such as finding resources and building a camp are still difficult for the model.
SIMA is not intended to replace AI-based NPCs, but to be a player. SIMA project co-director Frederic Bess said it was too early to say how such AI agents could benefit games outside of research.
